A predictor that ends up reasoning
A large language model is trained for one mechanical job: predict the next token. Nothing in that objective says 'reason.' Yet when you ask a strong model to plan a trip, debug code, or work through a logic puzzle it has never seen, it often produces a chain of sensible intermediate steps and lands on the right answer. We call this reasoning in LLMs — but the word is doing a lot of work, so let's be precise about what it does and doesn't mean.
The honest framing: an LLM does not run a logic engine. It generates text token by token, and reasoning is the behaviour that emerges when generating each next token requires combining facts, applying a rule, or tracking a quantity across several steps. The reasoning lives in the process of generation, not in a separate module bolted on the side.
A loop diagram of autoregressive generation: the model outputs a token, appends it to the input, and repeats to produce the next token.
Recall versus reasoning
Two very different skills hide under one polished answer. The first is world-knowledge recall: facts compressed into the weights during training — capital cities, famous theorems, common idioms. Ask 'what is the boiling point of water at sea level?' and the model is essentially looking it up from memory. The second is reasoning rather than memorization: deriving an answer the training data never stated outright, by chaining steps together.
The boundary is fuzzy and that is exactly why it matters. 'What is 17 times 24?' looks like reasoning, but a common product may simply be memorized. 'What is 4173 times 2891?' almost certainly was not in the training data, so a correct answer must come from a genuine multi-step procedure. A big part of judging a model's capabilities is telling these two apart — a theme we return to in the final guide.
Two speeds: intuition and deliberation
A useful borrowed metaphor is fast versus slow thinking — deliberate versus intuitive processing. When a model answers in a single token of output ('the answer is 42'), it is committing to a guess from its first impression, like a person blurting out an answer. When it writes out the working first and only then states the result, it gives itself room to correct course mid-stream.
Each next token is drawn from a distribution conditioned on every token so far, so the intermediate 'scratch-paper' tokens the model writes become context that shapes its final answer.
This is the single most important practical insight in the whole field: *for hard problems, letting the model produce more intermediate tokens before the answer usually raises accuracy.* The next three guides are essentially four ways to cash in on that insight — by prompt, by training, and by spending extra compute at answer time.
Skills nobody programmed in
Where do these abilities come from? Two sources. Some are emergent reasoning skills — capabilities that are weak or absent in small models and appear, sometimes sharply, only once a model crosses a certain scale of size and training. Others are picked up on the spot: in-context skill acquisition, where a couple of worked examples in the prompt teach the model a brand-new format or rule for the duration of that conversation, with no weight updates at all.
A log-log plot of loss decreasing as model scale increases, illustrating neural scaling laws.
Both are why an LLM feels less like a fixed program and more like a flexible substrate you can steer. It also explains why the same model can look brilliant on one phrasing and clueless on another: the reasoning is real but fragile, sensitive to how the question is set up. Keep that tension — powerful yet brittle — in mind; it is the through-line of this entire track.